Speech to Text FAQ

Is this software open-source?

No. It’s closed-source, proprietary software.

It does contain open-source components licensed under MIT, Apache, BSD, and other permissive licences. For more information about these components and their licences, please read "Pre-existing IP.txt" in the software’s installation folder.

Additionally, after releasing version 1.0, I open sourced some parts of the server that runs this website.

Is this software an AI slop?

No. All source code I compile and ship in the installer, or run on this server, is written by me personally, except third-party dependencies, nearly all of which are open-source software.

I have only used LLMs for the following tasks.

To review the code I wrote.
To generate fragments of HTML or CSS markup.
To generate some functions in the code I don’t ship, like build automation or QA infrastructure. This includes Python functions in the ErrorRateMethod.pdf document.
To review and edit the text in the GUI, on this web site, and in the documentation.

Does the software send my audio anywhere?

No, the speech is transcribed locally on your computer.

The only components of the software which use internet:

“Update” button contacts this server to query the latest published version of the software.
Licence helper tool contacts this server to obtain software licences.
Model downloader tool contacts this server, as well as other computers on the internets, to obtain speech to text models.
Crash uploader tool contacts this server to upload crash reports, after you click the “Upload” button.

See privacy policy for more information.

What hardware do I actually need?

I don’t know because I have limited selection of hardware to test. Make sure to evaluate the software on your specific machine before purchasing. I provide 30 days evaluation licences partly for that reason.

This answer was written for the version 1.0 of the software. And it assumes you download and use the 3.4 GB Voxtral Mini BCML speech to text model, which I recommend – see the performance page for more information.

If you have a discrete GPU launched within the last 10 years equipped with at least 4 GB VRAM, the software will probably work fine.

For integrated graphics it’s more nuanced. My best estimate, you need at least 3 TFLOPS FP16 compute, at least 8 GB memory, and at least 50 GB/second memory bandwidth.
For example, Ryzen 5 5600U iGPU in a laptop with dual-channel DDR4-3200 memory is OK. OTOH, a desktop with a faster Ryzen 7 5700G iGPU but slower dual-channel DDR4-2400 doesn’t reach realtime rate. The memory bandwidth in the desktop is insufficient, 38.4 GB/second.

What’s the difference between personal and commercial licence types?

Two differences.

As written in the terms of service, a personal licence is permitted for non-commercial use only.
When using a personal licence, window titles of the transcribe applications say so.

Licence type does not affect anything else.

What happens if I activate a permanent licence, then reinstall the OS?

Your old licence remains valid. Launch LicenceHelper.exe, sign in with e-mail and password, click the download button. Also, “Issued licences” table on the account page has a download link.

What happens if I activate a permanent licence, then upgrade or replace the computer?

Depending on what specifically you have upgraded or replaced, your old licence may become invalid.

If you activated the licence more than 12 months ago, you should see a transfer link on the account page.

If it’s been less than 12 months, the automatic transfer isn’t available yet. Please contact support from the e-mail you have specified when creating your user account, and I’ll look into issuing a new licence for your new hardware.

What happens after an evaluation licence expires?

Legally, as I wrote on the Terms of Service page, you may only use the free evaluation versions for 30 days on any particular computer you own. The software is not a shareware.

Technically, when an evaluation licence expires, the software will stop running until you either purchase a licence or obtain a new evaluation licence.

I reserve the right to adjust how evaluation licences are issued and tracked. Still, the copy protection is designed to allow at least 30 days of free evaluation per computer. If it doesn’t work for you, please contact support.

What should I do if the software doesn’t work?

I strongly encourage all customers to evaluate the software fully before buying a licence. The backend doesn’t depend on licence type. The only thing copy protection does is preventing the software from launching without a valid licence.

If you have a computer which matches or exceeds the recommended system requirements from the installation guide, and the software is broken due to a bug, please enable diagnostic log files in advanced preferences. Then e-mail that log file from your %LOCALAPPDATA%\Speech to Text\Logs. Ideally, include the system information saved by dxdiag.exe.

I don’t promise I will respond to every e-mail, nor fix every issue you report. Some classes of bugs are specific to hardware (especially the GPU hardware) or software environment, which means I might not be able to reproduce or fix.

What’s the difference between Voxtral and Whisper models?

Voxtral model is newer. It delivers much smaller latency, and in my tests is more reliable especially for low quality audio.

Whisper models require less resources. They support more languages. Whisper models can translate non-English speech to English text, if you check the “Translate to English” checkbox. Albeit I consider the translation feature experimental, having tested it only lightly.

Any plans for a macOS version? Linux version?

I would like to support both, but I can’t promise that will ever happen, and I don’t have a timeline.

At the time of writing, I have not even launched the windows-only 1.0 version.

This document was last updated on 2026-09-11